Uncalibrated Near-Light Photometric Stereo

Photometric stereo (PS) [3] is a technique to accurately recover the normal map of a 3D scene from several pictures (at least three) taken from the same view point and under different illumination conditions. When the light directions and intensities are known, photometric stereo can be solved as a linear system. When the illumination is not known, one needs to solve a much harder problem: uncalibrated photometric stereo. Typical assumptions are the Lambertian reflectance, orthographic projection, absence of shadows and interreflections and that the light sources are far away from the object. In particular, the last assumption allows to consider parallel illumination and, consequently, a simpler image formation model. The distant light assumption is a reasonable approximation as long as the dimensions of the scene are much smaller than the distance of the light sources. However, this may not be the case in many practical scenarios such as endoscopy, cultural heritage, reconstruction of big indoor objects, underground and underwater navigation, or full human body 3D reconstruction. Motivated by this fact, we introduce for the fist time an uncalibrated near-light photometric stereo method where no prior information about light position and intensities is needed. Only in [1] uncalibrated near-lights were considered. However, the method only recovers depth cues obtained from particular illumination configurations (lights moving on a line or plane), while in our algorithm we consider illuminants distributed arbitrarily in front of the object. We achieve this by first analyzing the reconstruction ambiguities and then by introducing an iterative technique to solve for the normals, reflectance and lights. We demonstrate the practical use and accuracy of our algorithm with real world experiments and compare it with the state-of-art in uncalibrated distant light photometric stereo. The image formation model typically used for the near-light case under the Lambertian reflectance is